Abstract
In patients undergoing invasive mechanical ventilation (IMV), IVCdmax, TAPSE, and the tricuspid E-wave peak show strong correlations with VExUS grading. A right ventricular function assessment model based on VExUS grading and the classification and regression tree (CART) algorithm can effectively evaluate right ventricular performance and may serve as a useful tool in identifying venous congestion.
Introduction
In the intensive care unit (ICU), invasive mechanical ventilation (IMV) is a critical intervention that maintains respiration but simultaneously places significant stress on the right ventricle (RV) due to increased intrathoracic pressure. If the RV fails to pump efficiently, blood "backs up," leading to venous congestion, organ dysfunction (such as acute kidney injury), and increased mortality.
While echocardiography remains the gold standard for RV assessment, it is often difficult to perform or interpret in critically ill patients due to poor acoustic windows or the need for high-level expertise. The Venous Excess Ultrasound (VExUS) grading system—which evaluates the inferior vena cava (IVC) and Doppler waveforms in the liver, gut, and kidneys—has emerged as a powerful way to quantify this congestion. This study introduces a transparent, decision-tree-based model to assess RV function by linking bedside ultrasound indicators with VExUS grading, offering clinicians a practical roadmap for hemodynamic management.
Methods
The researchers conducted a cross-sectional study of patients undergoing IMV, collecting clinical markers (such as NT-proBNP and Central Venous Pressure) alongside detailed ultrasound measurements of RV structure and function. The core of the approach was the Classification and Regression Tree (CART) algorithm. Unlike "black-box" deep learning models, CART produces a visible flowchart (a decision tree) that identifies which variables provide the most "information gain" to accurately categorize patients into VExUS grades (0 to 3).
To optimize the model, the data was split into an 8:2 ratio for training and validation. The algorithm ranked variables by importance to determine the most efficient diagnostic path.
Study Population and Datasets
The study utilized a primary clinical dataset collected at the Harrison International Peace Hospital (Hengshui, China) between October 2023 and December 2024.
As shown in the flowchart above, the study initially screened 127 patients. After applying strict exclusion criteria—such as obesity, short-term ventilation, or pre-existing organ failure—to reduce confounding factors, 80 patients were included for final data collection. These patients were categorized into VExUS 0 (no congestion, n=26) and VExUS ≥1 (presence of congestion, n=54).
Key Results and Insights
The CART model identified IVCdmax (maximum diameter of the IVC) as the most critical "root node." If the IVC diameter exceeds 1.975 cm, the model immediately flags a higher risk of congestion. Beyond volume status, the model incorporates TAPSE (a measure of RV contraction) and the tricuspid E-wave peak (a marker of diastolic filling) to refine the diagnosis.
Key performance metrics of the model include:
- Model Accuracy: The CART model achieved an overall accuracy of 87.5% on the test set. Optimization of the tree depth (as seen in Figure 2) allowed the model to reach high training accuracy before plateauing.
- Discriminatory Power: The Area Under the Curve (AUC) was excellent: 0.917 for Grade 0, 0.828 for Grade 1, and a perfect 1.0 for identifying moderate-to-severe congestion (Grades 2 and 3).
- Correlation Strengths: IVCdmax showed a strong positive correlation with VExUS grading (r = 0.773), while TAPSE (r = -0.670) and tricuspid E-wave peak (r = -0.648) were significantly negatively correlated.
- Biomarker Validation: Higher VExUS grades correlated strongly with stress biomarkers like NT-proBNP and Cardiac Troponin I (cTnI), validating that venous congestion is a systemic signal of myocardial distress.
Discussion and Limitations
The authors chose the CART algorithm specifically for its clinical transparency. In high-stakes ICU environments, a model that explains why it reached a conclusion is often more valuable than opaque neural networks. However, several hurdles remain for widespread adoption:
- Sample Size: With 80 patients, the model requires validation in larger, multicenter cohorts to avoid overfitting.
- Operator Proficiency: All ultrasounds were performed by one experienced sonographer; real-world accuracy may vary among different clinicians.
- Dynamic Assessment: The study was cross-sectional. Future research should track how VExUS grades change over time in response to treatments like fluid removal or changes in ventilator settings.
- Ventilator Parameters: Technical parameters like driving pressure were not included in the model, which could provide further context for RV strain.
Summary updated